Fair use isn't a feature of copyright in every juristiction, which could make this a less than useful idea trying to create a global corpus of speech data.